Online Label Shift

OnlineLabelShift is the official implementation of the accompanying paper Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms. For more details, please refer to the paper.

Setup Environment

conda env update --file environment.yml

Run Online Label Shift Experiment

The following command runs the online label shift experiment. It expects the base model has been trained and saved under /model

python scripts/run_ols.py -d synthetic -m logreg --do-all 1 -t 1000 --save 1

To see all the options

python scripts/run_ols.py -h

Train model

This script supports model training for synthetic, cifar10, and mnist datasets.

python scripts/train_model.py -d <data> -m <model> -e <epoch>

The corresponding models are:

Data	Model
synthetic	logreg
mnist	fcn
cifar10	resnet18

Generate Synthetic Data

To run experiments on synthetic data, one should first generate the data with the following command:

python scripts/gen_synth_data.py